Title: N' Okaeme, P' Zanchetta, M' Sumner PEMC group University of Nottingham, UK
1N. Okaeme, P. Zanchetta, M. Sumner PEMC group
University of Nottingham, UK
ECON2 Marie Curie Mini-Conference July 9, 2008,
School of EEE, University of Nottingham, UK
Automated Online Design of Robust Speed Digital
Controllers For Variable Speed Drives
2Outline
- Genetic Algorithm
- Digital Controller
- Mechanical Loads Investigated
- Experimental Approach
- Theoretical Approach
- Results
- Conclusion
3Aim of the research
Traditional controllers such as Proportional plus
Integral (PI), which are widely used in
industrial drives, may not give satisfactory
results in all operative conditions, above all in
the presence of largely variable loads.
Keep straightforward control implementation
Provide excellent performance and robustness to
variable loads
4Aim of the research
To investigate a control design
procedure Robustness to variable loads and
different load types Traditional discrete linear
control implementation (z domain
controllers) (Most likely need higher order
controllers) No need for a fixed control
structure Automated design No need for plant
modelling
On-line optimization procedure for experimental
control system design using Genetic Algorithm (GA)
5Aim of the research
6Outline
- Genetic Algorithm
- Digital Controller
- Mechanical Loads Investigated
- Experimental Approach
- Theoretical Approach
- Results
- Conclusion
7Genetic Algorithm
Stochastic global search method based on
biological evolution
GA
- Developed as a Part of Evolutionary Computing
introduced in 1960s By I. Rechenberg - GA proposed by John Holland in the mid-1970s
- Random Numerical Search and Optimisation
Technique that operates on a population of
potential solutions, termed individuals, applying
the principle of evolution, simulated by means of
mathematical operations that mimic the process of
selection, crossover and mutation. - A fitness function measures the fitness of an
individual to survive in a population of
individuals.
8Genetic Algorithm
Randomly Generate Initial Population
Yes
No
9Genetic Algorithm
Fitness function evaluation
- The quality of the system response to each load
operative condition is quantified online using
the Fitness function. - The fitness function is defined as a linear
function of the target specifications overshoot
(OS), rise time (tr), steady state error (ess),
steady state ripple (rss), system bandwidth (BW)
- Further requirements can be added - These indices are weighted individually and are
summed up to give a fitness value. The overall
quality of the closed loop response under
different load conditions is obtained by
weighting and then summing all of the relative
fitness values - FF (k1OS k2 tr k3 ess k4 rss k5
BW)JJnom - (k1OS k2 tr k3 ess k4 rss k5
BW)J2Jnom
10Genetic Algorithm
Selection
- Process that chooses the fittest individuals from
a population to continue into the next
generation. Principle of Survival of the
fittest - Proportionate selection the probability that an
individual advances to the next generation its
proportional to its relative fitness - Expected offsprings of an individual product of
the individuals relative fitness times the
number of individuals in the population.
11Genetic Algorithm
Crossover
- Crossover generates new individuals by exchanging
genetic material between individuals - Individual coding uses real number representation
- A random number between 0 and 1 is generated. If
it is smaller than a fixed crossover probability
then two individuals are randomly chosen and
crossed. The resulting offspring replaces the
parents in the new population. - C1 and C2 parents individuals C1new and C2new
offsprings - C1new ? C1(1-?) C2
- C2new ? C2(1-?) C1
0lt ?lt1 Crossover parameter
12Genetic Algorithm
Mutation
- Mutation effects a random variation upon the gene
of an individual with a fixed mutation
probability - A random number between 0 and 1 is generated. If
it is smaller than the mutation probability then
a gene is randomly chosen and mutated. The
resulting offspring replaces the parents in the
new population - Uniform Mutation
- Ci Ci1,., Cij ., Cin
Individual - Cinew Ci1,., Cijnew ., Cin
New individual after mutation - where Cijnew is a random value (with uniform
probability distribution) within the domain of
Cij
13Outline
- Genetic Algorithm
- Digital Controller
- Mechanical Loads Investigated
- Experimental Approach
- Theoretical Approach
- Results
- Conclusion
14Digital Controller
Encoding unstructured controllers
e(k)
u(k)
F1
F2
F3
F4
F5
Chromosome of a single controller
15Outline
- Genetic Algorithm
- Digital Controller
- Mechanical Loads Investigated
- Experimental Approach
- Theoretical Approach
- Results
- Conclusion
16Mechanical Loads Investigated
- Stiff shaft mechanical load
- J1 Jem 5J1 and B Bem 5B
- Flexible shaft mechanical load
-
-
- J2 JLem 5J2, D Dem 5D,
backlash1.5rad/sec
17Outline
- Genetic Algorithm
- Digital Controller
- Mechanical Loads Investigated
- Experimental Approach
- Theoretical Approach
- Results
- Conclusion
18Experimental Approach
Two identical permanent magnet DC servomotors
coupled along the same shaft. One serves as the
Driving motor and the other as the load motor.
19Experimental Approach
Power conversion stages DC to DC converter based
on a MOSFET H-Bridge configuration switching at
20kHz with nominal current of 5A Control xPC
target toolbox in Matlab-Simulink and interfaced
with the motors using the National Instrument
PCI-MIO-16XE-10 I/O board
20Experimental Approach
- GA based software that automatically designs
optimised digital regulators for robust control
of a permanent magnet DC variable speed drive has
been developed - Software applicable for control optimisation of
power electronics and drives systems in general - Optimisation of structure and parameters of
speed controllers online, while the drive is
being subject to variable mechanical loads. - Different mechanical loads in the experimental
control design tests are achieved by using a
programmable load emulator.
21Experimental Approach
- High Flexibility Original Software tool in Matlab
- - every type of application and control
structure - - fitness function adjustable to specs
requirements - The software defines the more appropriate order
and the best structure of the controller and
determines the optimum values of its parameters - Structure can include
- gain, pure integrator, PI regulator, real poles
and zeros, complex poles and zeros - The user can set
- - bounds for the parameters values
- - probability of mutation and crossover
- - guidance on the controller structure
22Summary on Experimental approach
Robust control design for loads with variable
inertia
- The GA software is written in Matlab language and
runs on the host PC - Each individual is then tested experimentally
online on the actual real rig - The speed closed loop dynamic for the range of
different mechanical loads is measured
Host PC runs GA in Matlab
23Summary on Experimental approach
- Optimization time
- Optimisations are performed through 30
generations with each generation having 40
individuals Each experiment takes approximately
5s. Total time for the optimisation is
approximately two hours - Experience has shown that improvements obtained
with longer optimizations are not substantial for
the sake of control robustness.
- Protections
- During the experiments, it is necessary to
prematurely stop the experimental test of badly
performing individuals (to reduce the time
duration of the experiments, as well as the risks
of damages to the hardware). - Suitable logical protection schemes both in the
control software and hardware implementation.
24Outline
- Genetic Algorithm
- Digital Controller
- Mechanical Loads Investigated
- Experimental Approach
- Theoretical Approach
- Results
- Conclusion
25Background
26(No Transcript)
27Identification of Nominal Model
- Offline simulation model
- Sub-optimal controller
- Model of mechanical loads
- Achieved using GA
- Run of 50 generations 50 individuals
- Identified over drives speed range
- Parameters identified at particular speeds
- Range of operation 0 3000 rpm
- Parameters values vs. speed plots
- Nominal model average of parameters
28Identification of Nominal Model
- Mechanical load with stiff shaft
Variation of the friction of the nominal load
Variation of the Inertia of the nominal load
29Identification of Nominal Model
- Mechanical load with stiff shaft
GA identified nominal model response matched with
experimental data reference speed 105 rad/s
GA identified nominal model response matched with
experimental data reference speed 210 rad/s
30Identification of Nominal Model
- Mechanical load with flexible shaft
Variation of the Inertia of the nominal load
Variation of Damping coefficient of nominal load
with speed
31Identification of Nominal Model
- Mechanical load with flexible shaft
Backlash causes oscillations reference speed
105 rad/s
reference speed 210 rad/s
32Plant Uncertainty Model
- Feedback Uncertainty Model for Gp is selected
- Models Unstructured Uncertainty
- Poles crossing from left right-half plane
33Plant Uncertainty Model
- Define weighting function, W2
Load with flexible Shaft W2 0.002(0.004s 1)
Load with Stiff Shaft W2 0.00075(0.325s 1)
34Performance criteria
- Robust stability condition
- Where S is
- and C is the controller
transfer function
35Closed loop control design
- Optimization implemented using GA
- Offline model of experimental system
- Fitness function
- Nominal performance Robust Stability
36Outline
- Genetic Algorithm
- Digital Controller
- Mechanical Loads Investigated
- Experimental Approach
- Theoretical Approach
- Results
- Conclusion
37Results
- GA designed controller response
Load with Stiff Shaft Inertia 5J, Friction
5B Tr 100ms, Ts 110ms
Load with Stiff Shaft Inertia J, Friction
5B Tr 50ms, Ts 55ms
38Results
- GA designed controller response
Load with flexible shaft with backlash Inertia
5J, Damping 5D Tr 60ms, Ts 65ms
Load with flexible shaft with backlash Inertia
J, Damping 5D Tr 52ms, Ts 57ms
39Outline
- Genetic Algorithm
- Digital Controller
- Mechanical Loads Investigated
- Experimental Approach
- Theoretical Approach
- Results
- Conclusion
40Conclusion
- Simple effective approach
- Reduces commissioning time for drives
- Requires little user interaction
- Experimental robust control design
- Does not require modelling
- Optimization implemented directly on rig
- Theoretical robust control design
- Faster implementation within computer
- Less stress experienced by load machine